Change secondary network Pod controller to subscribe to CNIServer events #5767
Conversation
jianjuns
added
the
area/secondary-network
jianjuns
force-pushed
the
secnet-store
branch
2 times, most recently
from
059dc34 to
24ece85
Compare
| pc.podCache.SetPodCNIDeleted(containerInfo) | ||
| } | ||
| } | ||
| pc.queue.Add(podKeyGet(event.PodNamespace, event.PodName)) |
There was a problem hiding this comment.
I guess this is a key code difference in this patch: with CNI events being handled asynchronously, we need to ensure we enqueue the Pods
There was a problem hiding this comment.
Yes. Actually I am thinking to change the Pod event handling part in the next PR, so we can create/delete secondary interfaces right after CNI add/delete.
jianjuns
changed the title
jianjuns
force-pushed
the
secnet-store
branch
from
24ece85 to
06bc160
Compare
cmd/antrea-agent/agent.go
Outdated
| if err != nil { | ||
| return fmt.Errorf("error initializing CNI server with cniPodInfoStore cache: %v", err) | ||
| } | ||
| } else { | ||
| err = cniServer.Initialize(ovsBridgeClient, ofClient, ifaceStore, podUpdateChannel, nil) | ||
| err = cniServer.Initialize(ovsBridgeClient, ofClient, ifaceStore, podUpdateChannel) |
There was a problem hiding this comment.
we should merge the two branches since cniServer is decoupled with SecondaryNetwork
| ContainerID: containerConfig.ContainerID, | ||
| NetNS: netNS, |
There was a problem hiding this comment.
I wonder how it handles restart case if it relies on the netNS which is not persistent anywhere. There may be two cases it can't handle:
- If agent is restarted after CNI event is handled but before PodController handles it, it can't create no longer create the secondary interface.
- After agent is restarted, it can't clean up the secondary interface as the cache is always empty?
Should it cache netNS in containerConfig and use containerConfig to initialize podCache?
There was a problem hiding this comment.
Right. Today we do not persist any secondary interface configuration, and I plan to implement that in the next PR. We can definitely save net NS with the primary interface OVS port and I thought about that too. But I want to mention that only address the case antrea-agent restarts after Pod's primary interfaces are created but before any secondary interface is created, as only secondary interface creation (not deletion) will require net NS. I am still thinking is it worthwhile to cover that corner case or not.
The commit moves CNI Pod information store from CNIServer to secondary network PodController, and let PodController subscribe CNIServer Pod events to maintain the Pod CNI information in the store. This way simplifies CNIServer and removes secondary network logic from it. Signed-off-by: Jianjun Shen <shenj@vmware.com>
jianjuns
force-pushed
the
secnet-store
branch
from
06bc160 to
33942ac
Compare
|
/test-all |
The commit moves CNI Pod information store from CNIServer to secondary network PodController, and let PodController subscribe to CNIServer Pod events to maintain the Pod CNI information in the store. This way simplifies CNIServer and removes secondary network logic from it.
Issue: #5278